融合多模态数据的自闭症谱系障碍儿童智能化识别及其有效性  

Intelligent identification of children with autism spectrum disorder integrating multimodal data and its effectiveness

作　　者：廖梦怡 陈靓影[1] 王广帅[3] 彭世新 Mengyi Liao;Jingying Chen;Guangshuai Wang;Shixin Peng(National Engineering Research Center for E-Learning,Central China Normal University,Wuhan 430079,China;College of Computer Science and Technology,Pingdingshan University,Pingdingshan 467000,China;College of Computer Science and Technology,Wuhan University,Wuhan 430072,China)

机构地区：[1]华中师范大学国家数字化学习工程技术研究中心,武汉430079 [2]平顶山学院计算机学院,平顶山467000 [3]武汉大学计算机学院,武汉430072

出　　处：《科学通报》2021年第20期2618-2628,共11页Chinese Science Bulletin

基　　金：国家自然科学基金(61977027);湖北省技术创新专项重大项目(2019AAA044);教育部人文社会科学项目(19YJC880068)资助。

摘　　要：自闭症发病率呈逐年上升趋势,早期发现和及时干预可以显著改善预后,能够很大程度上改善自闭症儿童的语言能力、认知能力以及行为习惯,因此,自闭症的早期识别工作意义重大.传统识别方法能够获得较好的识别结果,但过程耗时且高度依赖于专业人员的操作,而已有的智能化识别方法的识别精度难以满足应用需求.本研究探索了融合多模态数据的自闭症儿童智能化识别方法,通过对行为数据和认知数据的分析发现,自闭症儿童和典型发展儿童在眼动、面部表情、认知得分和认知反应时数据上存在显著性差异.本研究利用数据差异性分析进行特征选择,构建了融合多模态数据的自闭症儿童识别框架,该框架根据数据来源和时间同步性将数据进行分层融合,进而得到最终的识别结果.同时,将融合多模态数据的自闭症儿童智能化识别方法以及各单模态识别方法分别与传统方法识别结果进行一致性检验,验证融合多模态数据识别方法的泛化能力和有效性.结果表明,与各单模态识别方法相比,融合多模态数据的自闭症谱系障碍儿童识别与传统方法识别结果一致性程度最高,识别正确率与传统识别方法的正确率最接近,是一种有效的自闭症儿童智能化识别方法.The incidence of autism spectrum disorder(ASD)keeps increasing in recent years.Early identification and timely intervention can significantly improve the prognosis,and can greatly improve the language ability,cognitive ability and behavioral habits of children with ASD.Therefore,the early identification of children with ASD is of great significance.The traditional methods can get better identification results,but its process is time-consuming and highly dependent on experts.Most of the existing intelligent identification methods are based on single modal data and ignore the high-level information provided by multi modal fusion data,and their identification accuracy is difficult to satisfy the application.More accurate and rapid intelligent identification method is in badly needed to support clinical practice.In this study,we propose an intelligent identification method of children with ASD.Through the analysis of behavioral and cognitive data,it is found that there are significant differences in the data of eye movement,facial expression,cognitive performance and cognitive response time between children with ASD and typically developing children.Feature selection is carried out through data difference analysis.Then,a multimodal identification framework is constructed,which fuses the multimodal data in two layers according to the data source and time synchronization.In the first layer,the synchronous data from different sources are fused to form behavioral feature vector and cognitive feature vector,which are respectively sent into classifiers for decision.In the second layer,asynchronous data are fused,and each decision is weighted and fused by decision weight,and then the final recognition result is obtained.At the same time,the intelligent identification method and each single-modal identification method are respectively tested for consistency with the identification results of traditional method,to verify the generalization ability and effectiveness of the multimodal data fusion identification.The results indi

关 键 词：自闭症谱系障碍 智能化识别 单模态数据 多模态数据融合 一致性检验 

分 类 号：R749.94[医药卫生—神经病学与精神病学]